Which AI actually writes Indonesian well?

Two AI models answer the same Indonesian question, whether slang, a regional dialect, or formal register. You read both answers with the model names hidden, and pick the one that fits best. Only after your vote do we reveal which model wrote which answer, along with what each model's record in the arena looks like so far.

Because a visible name changes what is being judged. Shown the brand first, people rate the brand; shown only the writing, they rate the writing. Blind comparison is the only version of this that measures language quality rather than reputation. Which side an answer appears on is also decided by the server and randomised, so a habit of picking the left box does not quietly become a result.

No. Every answer is written ahead of time and stored, then served as a pair. That is a deliberate choice: it means the arena responds instantly, the same answer can be compared against several opponents without being rewritten, and what you judge is a fixed text rather than a fresh roll of the dice.

Entirely from the votes cast here. A Bradley-Terry model turns head-to-head outcomes into a strength estimate for each model, which is not the same as counting wins: beating an opponent that rarely loses moves the estimate far more than beating one that usually does, so no model is rewarded for happening to face easy opponents. A rank is only published once a model has been judged often enough for its interval to be meaningful; below that, an order would be noise dressed as a result.

Because with a limited number of votes, several models cannot be told apart. Each estimate carries an interval, and where two intervals overlap there is no evidence that one model sits above the other. Printing a single position for both would be a claim the votes do not support. A model whose interval is not yet meaningful is given the full range, which is the correct answer to "what rank is it" when the answer is not known.

Because forcing a choice between two answers that are both acceptable injects noise into the ranking. In categories judged on speaker acceptance rather than a prescriptive rule, two different answers being equally right is a normal situation, not an edge case.

A share of pairs are calibration items whose correct answer we already know. They measure how reliable the voting is overall, so they are deliberately excluded from the ranking itself. Votes cast in browsing mode, and votes returned too quickly to be a real reading, are also recorded but kept out of the ranking.

That bias is real and well documented in this kind of arena, which is why we correct for it. Alongside the raw rating we estimate a style-adjusted one, which accounts for answer length and formatting. The gap between the two is itself a finding: it shows how much of a preference was rhetoric rather than language competence.